Automated Vision-Based Sensing of Paediatric Pain Using 2D-Facial Landmark Trajectories Derived from 3D Geometric Normalisation and a Spatial-Temporal Attention Long Short-Term Memory (STA-LSTM) Network

Abstract

Background: Accurate pain assessment in children remains challenging because conventional behavioural assessment tools are subjective and provide only intermittent observations. This study evaluated a vision-based sensing system for automated paediatric pain detection using two-dimensional facial landmark trajectories derived from three-dimensional geometric normalisation of frontal facial video recordings. Methods: In a prospective observational study, 125 children aged 6–17 years undergoing elective surgery had frontal facial videos recorded pre- and postoperatively using a custom-developed iOS mobile application. Video frames were processed using the MediaPipe Face Mesh framework to estimate 478 facial landmarks with three-dimensional coordinates. After three-dimensional normalisation, the normalised x- and y-coordinates were segmented into one-second trajectories, and analysed using a previously developed Spatial-Temporal Attention Long Short-Term Memory (STA-LSTM) framework. Reference pain labels (pain vs. no pain) were derived from observer-rated revised Face, Legs, Activity, Cry, Consolability (r-FLACC) scores. Results: The sensing system analysed 971 annotated clips and was trained on 6438 balanced landmark trajectories. On an independent validation dataset, it achieved an area under the receiver operating characteristic curve of 0.994, with an accuracy of 97.15%, precision of 86.78%, recall of 97.83%, and F1-score of 91.97% for binary pain classification. Conclusions: A landmark-based, non-contact vision sensor combined with a STA-LSTM network achieved high performance for automated detection of observer-labelled pain states in children, supporting facial landmark trajectories as a practical signal representation for privacy-conscious continuous pain monitoring.